Your dark theme needs a daylight review
A page can look different after sunset without becoming easier to use. Check its text and controls under the lighting conditions readers will encounter.

A dark page that looks polished at night can be difficult to read beside a bright window. The design file may be unchanged. The viewing conditions have changed, and the display may have adjusted itself too. Reviewing the page under both conditions gives a better basis for a decision than choosing whichever screenshot looks more attractive.
The temptation is to explain the whole effect through eyesight. That goes beyond what a casual observation can establish. A glowing display in a dim room is not equivalent to an object seen under very low light. Broad claims about nighttime color perception cannot simply be applied to every interface on that display.
There are more immediate variables to check. Light reflected from the screen can obscure detail. Automatic brightness can change the display's output. Color adaptation can alter its appearance too: Apple documents True Tone as adjusting display color and intensity to suit ambient light. A comparison made with different settings can mix a theme change with a display change.
Compare the same task under known conditions
Begin with a page that contains the text people actually need to read. A large heading on an empty background is an easy case. An article's secondary links or a form's helper text may reveal a different problem. The review should include those elements at the size used in the shipped interface.
Choose a real task, such as finding an article's date and following a reference link. Keep that task constant when moving between lighting conditions. If the reader can locate the link in a dim room but loses it beside the window, the observation gives the designer a specific element to inspect.
Record the relevant display settings before making color changes. For a controlled comparison, hold the settings constant so the theme is the variable. Then perform an additional check with the automatic settings a reader may normally use. These checks answer different questions and should retain their own notes.
A screenshot is useful for reviewing layout and spacing. It cannot reproduce light reflecting from the original screen. Sending the screenshot to another display changes the viewing conditions again. Use screenshots to document the interface, with a short description of the environment in which an issue appeared.
Measure text before adjusting the palette
The WCAG 2.2 contrast guidance calls for at least 4.5:1 contrast for normal text under Success Criterion 1.4.3, with 3:1 for qualifying large text. Exceptions apply, including logos and incidental text. These are defined accessibility requirements for the text-background pair, not a score for how pleasant an entire theme feels.
Measure the colors where text is actually drawn. A translucent card over an image can produce different effective backgrounds at different positions. Checking only the nominal card color can miss the lower-contrast region beneath a line of text. Inspect the rendered combination, including a hover state if it changes the text color.
Passing the ratio does not resolve every reading difficulty. Text can still be too small for the task, or a line can be awkwardly long. A page can meet the ratio while placing so many secondary elements beside the body that the reader struggles to keep their place. Treat the measured result as one part of the review.
This also prevents a common overcorrection. Increasing saturation may make an accent seem livelier without improving the readability of muted paragraph text. When the problem concerns a date label, change that label's foreground and effective background first. Evaluate the surrounding palette after the required information is readable.
A theme preference does not measure reading performance
A reader may prefer the appearance of light text on a dark surface. That preference is real, but it answers a different question from proofreading accuracy. In Buchner and Baumgartner's 2007 experiments, proofreading performance favored dark text on a light background under the tested conditions, across both dark and typical office illumination.
That result should remain tied to the task and conditions studied. It does not establish that every reader should use a light theme for every activity. It does give a designer reason to test reading performance instead of inferring it from the appearance of a nighttime screenshot.
Where a product offers theme choice, review the same required content in each supported theme. A toggle creates a second interface state that needs maintenance. A later change to an alert color can leave one theme readable and damage the other if the review covers only the designer's preferred setting.
For the next release, capture one issue found in daylight and trace it to a particular text-background pair. Fix that pair and recheck the contrast. Then repeat the reading task. Keep the result beside the dim-room review so the final palette is supported by the conditions people will actually use.